oamazonasgabriel/ lfm2-1.2b-tool

39 2 weeks ago

LFM2-1.2B-Tool — a specialized 1.2B tool-calling model from Liquid AI, fine-tuned exclusively for concise and precise function calling. Uses a hybrid LIV+GQA architecture, outperforms thinking models of similar size without chain-of-thought overhead. 

ollama run oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU

Models

View all →

1 model

lfm2-1.2b-tool:f16-8gbGPU

2.3GB · 125K context window · Text · 2 weeks ago

Readme

LFM2-1.2B-Tool — Ollama Model (8-16 GB GPU) Tool-Calling Specialized

Model: oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU

Ollama Built by impacte.tech

DESCRIPTION

LFM2-1.2B-Tool is a specialized variant of Liquid AI’s LFM2 architecture, fine-tuned exclusively for concise and precise tool calling. It combines 10 double-gated Linear-Invariant-Valued (LIV) convolution blocks with 6 full-attention Grouped-Query Attention (GQA) blocks — a novel hybrid design that achieves 2-3x better perplexity per parameter than dense transformers. The model is designed to outperform similarly sized thinking models for function-calling tasks without any internal chain-of-thought overhead, delivering instant tool call generation for latency-sensitive applications. This Ollama packaging uses the official F16 GGUF quantization (~2.5 GB).

“Built to Run Anywhere.” — Liquid AI

The ultimate tool-calling companion for: 8 GB laptops · 16 GB desktops · Edge servers · Real-time assistants · IoT devices · API orchestration

Key Features

  • Purpose-built for tool calling: outperforms thinking models of similar size on function-calling benchmarks
  • No chain-of-thought overhead: instant tool call generation — ideal for latency-critical applications
  • Greedy decoding (temp=0): deterministic output guarantees syntactically valid function calls every time
  • Full precision: F16 weights preserve 100% model quality (zero quantization loss)
  • 128K native context (configured to 32K for 8 GB, expandable to 128K on 16 GB+)
  • Pythonic function calls: native format uses Python syntax func(arg="value") — natural and intuitive
  • Ollama API compatible: OpenAI-compatible tool schemas are automatically translated to native format
  • Multi-turn support: handles complex tool chains across multiple conversation turns
  • 9 languages: EN, AR, ZH, FR, DE, JA, KO, PT, ES
  • LFM Open License v1.0: royalty-free for non-commercial/research and small business (<$10M annual revenue)

Architecture

Property Value
Architecture Hybrid: 10 conv blocks + 6 full_attention GQA blocks
Parameters 1.17B — dense (all active per token)
Layers 16 (10 conv + 6 full_attention)
Hidden Dim 2048
Intermediate Dim 12288
Attention Heads 32
Key-Value Heads 8 (GQA, ratio 4:1)
Native Context 128,000 tokens
Configured Context 32,768 tokens (8-16 GB GPU optimized)
Vocabulary 65,536 tokens
Base Model LFM2-1.2B
Knowledge Cutoff Mid-2024
Modalities Text only
Chat Format ChatML-like (`<
Tool List Tokens <|tool_list_start|>, `<
Tool Call Tokens <|tool_call_start|>, `<
Tool Response Tokens <|tool_response_start|>, `<
Tool Call Format Pythonic: function_name(arg="value")
FFN Activation SwiGLU
Position Encoding RoPE (theta=1,000,000)
Tie Embeddings Yes
Supported Languages English, Arabic, Chinese, French, German, Japanese, Korean, Portuguese, Spanish
Quantization F16 (maximum precision)
Model Size ~2.5 GB
License LFM Open License v1.0
Upstream LiquidAI/LFM2-1.2B-Tool-GGUF
Base Model LiquidAI/LFM2-1.2B-Tool

REQUIREMENTS

Resource Minimum (8 GB) Recommended (16 GB)
RAM 4 GB 8 GB+
GPU Memory 2 GB VRAM (optional) 4 GB+ VRAM (optional)
Disk Space 5 GB free 10 GB+ free
Ollama Version 0.30.6+ Latest

Platform support: - CPU-only — runs on any modern CPU (efficient hybrid architecture) - Any GPU (NVIDIA, AMD, Intel Arc, Apple Silicon) with 2 GB+ VRAM - Edge servers & IoT — purpose-built for edge deployment

💡 Why 8-16 GB RAM? The model is ~2.5 GB. An 8 GB machine has ~4.2 GB of headroom — enough to run alongside your other applications. On 16 GB, you have ~11 GB headroom for larger contexts.

QUICK START

1. Install Ollama

# macOS (Homebrew)
brew install ollama

# Linux
curl -fsSL https://ollama.com/install.sh | sh

# Windows
# Download from https://ollama.com/download

2. Pull & Run

# Pull the model (downloads ~2.5 GB)
ollama pull oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU

# Run interactively
ollama run oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU

USAGE

CLI

# Interactive chat (tool calling via API recommended)
ollama run oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU

# With system prompt providing tools
ollama run oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU \
  --system "You are a tool-calling assistant. Use the available tools to answer questions."

# Override context for 16 GB systems
ollama run oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU \
  --num-ctx 65536

REST API — Tool Calling (Primary Use Case)

# Single tool call
curl -s http://127.0.0.1:11434/api/chat \
  -d '{
    "model": "oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU",
    "messages": [
      {"role": "user", "content": "What is the weather in Boston?"}
    ],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "get_weather",
          "description": "Get the current weather for a city",
          "parameters": {
            "type": "object",
            "properties": {
              "location": {"type": "string", "description": "City name"}
            },
            "required": ["location"]
          }
        }
      }
    ],
    "options": {
      "temperature": 0.0
    }
  }'

Multi-Turn Tool Calling

# Turn 1: User asks a question
curl -s http://127.0.0.1:11434/api/chat \
  -d '{
    "model": "oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU",
    "messages": [
      {"role": "user", "content": "Find me a hotel in San Francisco for July 4-6 with a pool"}
    ],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "search_hotels",
          "description": "Search hotels by location, dates, and amenities",
          "parameters": {
            "type": "object",
            "properties": {
              "location": {"type": "string"},
              "check_in": {"type": "string"},
              "check_out": {"type": "string"},
              "amenities": {"type": "array", "items": {"type": "string"}}
            },
            "required": ["location", "check_in", "check_out"]
          }
        }
      }
    ],
    "options": {"temperature": 0.0}
  }'

# Then feed the tool response back to get the final answer
curl -s http://127.0.0.1:11434/api/chat \
  -d '{
    "model": "oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU",
    "messages": [
      {"role": "user", "content": "Find me a hotel in San Francisco for July 4-6 with a pool"},
      {"role": "assistant", "content": "",
       "tool_calls": [{"function": {"name": "search_hotels", "arguments": {"location": "San Francisco", "check_in": "2024-07-04", "check_out": "2024-07-06", "amenities": ["pool"]}}}]
      },
      {"role": "tool", "content": "{\"hotels\": [{\"name\": \"Marriott Marquis\", \"pool\": true, \"price\": 299}, {\"name\": \"Hilton Union Square\", \"pool\": true, \"price\": 249}]}"}
    ],
    "options": {"temperature": 0.0}
  }'

Basic Chat (No Tools)

curl -s http://127.0.0.1:11434/api/chat \
  -d '{
    "model": "oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU",
    "messages": [{"role": "user", "content": "Hello! What can you do?"}]
  }'

OpenAI-Compatible Endpoint

curl -s http://127.0.0.1:11434/v1/chat/completions \
  -d '{
    "model": "oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU",
    "messages": [{"role": "user", "content": "Get me the stock price for AAPL."}],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "get_stock_price",
          "description": "Get current stock price",
          "parameters": {
            "type": "object",
            "properties": {
              "symbol": {"type": "string"}
            },
            "required": ["symbol"]
          }
        }
      }
    ],
    "stream": false
  }'

Python (ollama library)

pip install ollama

import ollama

# Tool calling
response = ollama.chat(
    model='oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU',
    messages=[{'role': 'user', 'content': 'What is the weather in Boston?'}],
    tools=[{
        'type': 'function',
        'function': {
            'name': 'get_weather',
            'description': 'Get current weather',
            'parameters': {
                'type': 'object',
                'properties': {
                    'location': {'type': 'string', 'description': 'City name'}
                },
                'required': ['location']
            }
        }
    }],
)
print(response.message.tool_calls)

# Multi-turn with tool response
response = ollama.chat(
    model='oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU',
    messages=[
        {'role': 'user', 'content': 'What is the weather in Boston?'},
        {'role': 'assistant', 'content': '', 'tool_calls': [{'function': {'name': 'get_weather', 'arguments': {'location': 'Boston'}}}]},
        {'role': 'tool', 'content': '{"temperature": 72, "conditions": "sunny"}'},
    ],
)
print(response.message.content)

JavaScript (ollama.js)

npm install ollama

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU',
  messages: [{ role: 'user', content: 'Get the stock price for AAPL.' }],
  tools: [{
    type: 'function',
    function: {
      name: 'get_stock_price',
      description: 'Get current stock price',
      parameters: {
        type: 'object',
        properties: {
          symbol: { type: 'string' }
        },
        required: ['symbol']
      }
    }
  }],
})
console.log(response.message.tool_calls)

SAMPLING PARAMETERS

These are baked into the model via its Modelfile:

Parameter Value Rationale
num_ctx 32768 Generous context for multi-turn tool chains on 8-16 GB RAM
num_gpu 99 Offload all 16 layers to GPU
num_thread 6 Conservative for CPU-friendly operation
num_batch 256 Moderate batch size for 8 GB systems
temperature 0.0 Greedy decoding — official recommendation for deterministic tool calls
top_k 50 Safety net (no effect at temp=0)
top_p 0.9 No effect at temp=0
repeat_penalty 1.05 Light repetition penalty for natural flow
stop <\|im_start\|>, <\|im_end\|> Chat template tokens
stop <\|tool_list_start\|>, <\|tool_list_end\|> Tool list boundary tokens
stop <\|tool_call_start\|>, <\|tool_call_end\|> Tool call boundary tokens
stop <\|tool_response_start\|>, <\|tool_response_end\|> Tool response boundary tokens

⚠️ Important: Always use temperature=0.0 for production tool-calling pipelines. Higher temperatures may introduce syntax errors in Pythonic function calls.

Equivalent llama.cpp Command

./llama-cli \
  -m LFM2-1.2B-Tool-F16.gguf \
  -ngl 99 \
  -c 32768 \
  --temp 0.0 \
  --top-k 50 \
  --repeat-penalty 1.05 \
  -cnv

MEMORY & PERFORMANCE

RAM Budget

Component 8 GB System (32K ctx) 16 GB System (64K ctx)
Model weights (F16) ~2.5 GB ~2.5 GB
KV cache (FP16) ~1.0 GB ~2.0 GB
Ollama overhead ~0.3 GB ~0.3 GB
Total ~3.8 GB ~4.8 GB
Headroom ~4.2 GB ~11.2 GB

Context Window Scaling (8 GB, with q8_0 KV cache)

Context KV Cache Total Fits?
16,384 ~0.25 GB ~3.05 GB ✅ Lots of headroom
32,768 (default) ~0.50 GB ~3.30 GB ✅ Comfortable
65,536 ~1.00 GB ~3.80 GB ✅ Yes
131,072 (native) ~2.00 GB ~4.80 GB ✅ Yes (with headroom)

Performance

Hardware Tool Call Latency Text Generation
RTX 4060 (8 GB) ~10-30ms ~150-400 tok/s
RTX 3060 (12 GB) ~8-20ms ~200-500 tok/s
Apple Silicon M2 ~15-40ms ~100-300 tok/s
Intel i7 (CPU-only) ~50-200ms ~20-60 tok/s

TROUBLESHOOTING

Symptom Fix
Model not found Run ollama pull oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU first
Tool calling returns text instead of function call Ensure you’re passing tools array in /api/chat (not /api/generate)
Malformed function calls Make sure temperature is set to 0.0 — any randomness can break Pythonic syntax
Slow on CPU Normal for CPU; enable any GPU with num_gpu 99
OOM on 8 GB Reduce num_ctx to 16384 or enable q8_0 KV cache
Multi-turn not working Pass full conversation history including previous tool calls and responses

Optimizations for 8 GB Systems

# Set KV cache to q8_0 for ~50% memory reduction
export OLLAMA_KV_CACHE_TYPE=q8_0
ollama serve

# In another terminal:
ollama run oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU

# Or reduce context at runtime:
ollama run oamazonasgabriel/lfm2-1.2b-tool:f16-8gbGPU --num-ctx 16384

ALTERNATIVE MODELS

# Ultra-lightweight version of this family (4-8 GB)
ollama pull oamazonasgabriel/lfm2.5-230m:bf16-8gbGPU

# Larger general-purpose model (needs 16 GB)
ollama pull oamazonasgabriel/qwen3.5-9b:q4-16gbGPU

# Larger MoE model (needs 24 GB)
ollama pull oamazonasgabriel/qwen3.6-35b-a3b:q4-24gbGPU

# Same architecture, general instruct version
ollama pull hf.co/LiquidAI/LFM2.5-1.2B-Instruct-GGUF:Q4_K_M

# Quantized version of this model (smaller, faster)
ollama pull hf.co/LiquidAI/LFM2-1.2B-Tool-GGUF:Q4_K_M

CREDITS

This Ollama packaging brings together work from multiple teams and individuals:

Role Entity Contribution
Model Architecture & Training Liquid AI, Inc. Designed the hybrid LIV+GQA architecture, fine-tuned LFM2-1.2B for concise and precise tool calling without chain-of-thought. Based on the LFM2 architecture described in arXiv:2511.23404.
Original Model Liquid AI LFM2-1.2B-Tool tool-calling specialized checkpoint in native Transformers format.
GGUF Conversion Liquid AI / llama.cpp community Converted to GGUF format via llama.cpp for broad compatibility with Ollama, LM Studio, and other GGUF-based runtimes.
F16 Quantization Liquid AI Native F16 export preserving full model fidelity.
Ollama Modelfile & Packaging impacte.tech Ollama Modelfile configuration, greedy-decoding parameter tuning for 8-16 GB RAM targets, documentation, and registry publishing.
Benchmarking & Validation Liquid AI (official), impacte.tech (Ollama integration test) Verified on RTX 4060, RTX 3060, Apple Silicon, Intel CPU.
License Liquid AI, Inc. LFM Open License v1.0

Citation

If you use this model in your work, please cite the original LFM2 technical report:

@article{liquidai2025lfm2,
  title={LFM2 Technical Report},
  author={Liquid AI},
  journal={arXiv preprint arXiv:2511.23404},
  year={2025}
}

@misc{liquidAI2026tool,
  author = {Liquid AI},
  title = {LFM2-1.2B-Tool: Concise and Precise Tool Calling for Edge Devices},
  year = {2026},
  note = {https://huggingface.co/LiquidAI/LFM2-1.2B-Tool}
}

Acknowledgments

  • The llama.cpp project (github.com/ggerganov/llama.cpp) for the GGUF format and efficient inference engine that makes models like this accessible on commodity hardware.
  • The Ollama team (ollama.com) for the seamless model distribution platform.
  • Hugging Face (huggingface.co) for model hosting and the Hub ecosystem.

LINKS

Resource URL
This model on Ollama https://ollama.com/oamazonasgabriel/lfm2-1.2b-tool
Upstream model (HuggingFace) https://huggingface.co/LiquidAI/LFM2-1.2B-Tool
GGUF files (HuggingFace) https://huggingface.co/LiquidAI/LFM2-1.2B-Tool-GGUF
Liquid Nanos collection https://huggingface.co/collections/LiquidAI/liquid-nanos
LFM2 technical report https://arxiv.org/abs/2511.23404
Liquid AI blog https://www.liquid.ai/blog/introducing-liquid-nanos-frontier-grade-performance-on-everyday-devices
Liquid AI https://www.liquid.ai/
Ollama docs https://docs.ollama.com
Built by impacte.tech